Changes to partition breakout rules.

Changes to the breakout behavior for partition selection.
The biggest impact is on speed 0 where encode speed in
some cases more than doubles with typically less than 1%
impact on quality.

Speed 0 encode speed impact examples
Animation test clip: +128%
Park Joy:  +59%
Old town Cross: + 109%

Change-Id: I222720657e56cede1b2a5539096f788ffb2df3a1

vp9/encoder/vp9_encodeframe.c

@@ -2377,11 +2377,20 @@ static void rd_pick_partition(VP9_COMP *cpi, ThreadData *td,
                                bsize >= BLOCK_8X8;
   int partition_vert_allowed = !force_horz_split && xss <= yss &&
                                bsize >= BLOCK_8X8;
-  (void) *tp_orig;
+
+  int64_t dist_breakout_thr = cpi->sf.partition_search_breakout_dist_thr;
+  int rate_breakout_thr = cpi->sf.partition_search_breakout_rate_thr;
+
+  (void)*tp_orig;
 
   assert(num_8x8_blocks_wide_lookup[bsize] ==
              num_8x8_blocks_high_lookup[bsize]);
 
+  // Adjust dist breakout threshold according to the partition size.
+  dist_breakout_thr >>= 8 - (b_width_log2_lookup[bsize] +
+      b_height_log2_lookup[bsize]);
+  rate_breakout_thr *= num_pels_log2_lookup[bsize];
+
   vp9_rd_cost_init(&this_rdc);
   vp9_rd_cost_init(&sum_rdc);
   vp9_rd_cost_reset(&best_rdc);
@@ -2410,9 +2419,13 @@ static void rd_pick_partition(VP9_COMP *cpi, ThreadData *td,
                                 force_vert_split);
     do_split &= bsize > min_size;
   }
-  if (cpi->sf.use_square_partition_only) {
-    partition_horz_allowed &= force_horz_split;
-    partition_vert_allowed &= force_vert_split;
+
+  if (cpi->sf.use_square_partition_only &&
+      (bsize > cpi->sf.use_square_only_threshold)) {
+    if (!vp9_active_h_edge(cpi, mi_row, mi_step) || x->e_mbd.lossless)
+      partition_horz_allowed &= force_horz_split;
+    if (!vp9_active_v_edge(cpi, mi_row, mi_step) || x->e_mbd.lossless)
+      partition_vert_allowed &= force_vert_split;
   }
 
   save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
@@ -2489,27 +2502,17 @@ static void rd_pick_partition(VP9_COMP *cpi, ThreadData *td,
       }
 
       if (this_rdc.rdcost < best_rdc.rdcost) {
-        int64_t dist_breakout_thr = cpi->sf.partition_search_breakout_dist_thr;
-        int rate_breakout_thr = cpi->sf.partition_search_breakout_rate_thr;
-
         best_rdc = this_rdc;
         if (bsize >= BLOCK_8X8)
           pc_tree->partitioning = PARTITION_NONE;
 
-        // Adjust dist breakout threshold according to the partition size.
-        dist_breakout_thr >>= 8 - (b_width_log2_lookup[bsize] +
-            b_height_log2_lookup[bsize]);
-
-        rate_breakout_thr *= num_pels_log2_lookup[bsize];
-
         // If all y, u, v transform blocks in this partition are skippable, and
         // the dist & rate are within the thresholds, the partition search is
         // terminated for current branch of the partition search tree.
-        // The dist & rate thresholds are set to 0 at speed 0 to disable the
-        // early termination at that speed.
-        if (!x->e_mbd.lossless &&
-            (ctx->skippable && best_rdc.dist < dist_breakout_thr &&
-            best_rdc.rate < rate_breakout_thr)) {
+        if (!x->e_mbd.lossless && ctx->skippable  &&
+            ((best_rdc.dist < (dist_breakout_thr >> 2)) ||
+             (best_rdc.dist < dist_breakout_thr &&
+              best_rdc.rate < rate_breakout_thr))) {
           do_split = 0;
           do_rect = 0;
         }
@@ -2619,11 +2622,21 @@ static void rd_pick_partition(VP9_COMP *cpi, ThreadData *td,
       if (sum_rdc.rdcost < best_rdc.rdcost) {
         best_rdc = sum_rdc;
         pc_tree->partitioning = PARTITION_SPLIT;
+
+        // Rate and distortion based partition search termination clause.
+        if (!x->e_mbd.lossless &&
+            ((best_rdc.dist < (dist_breakout_thr >> 2)) ||
+             (best_rdc.dist < dist_breakout_thr &&
+              best_rdc.rate < rate_breakout_thr))) {
+          do_rect = 0;
+        }
       }
     } else {
       // skip rectangular partition test when larger block size
       // gives better rd cost
-      if (cpi->sf.less_rectangular_check)
+      if ((cpi->sf.less_rectangular_check) &&
+          ((bsize > cpi->sf.use_square_only_threshold) ||
+           (best_rdc.dist < dist_breakout_thr)))
         do_rect &= !partition_none_allowed;
     }
     restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
@@ -2632,7 +2645,7 @@ static void rd_pick_partition(VP9_COMP *cpi, ThreadData *td,
   // PARTITION_HORZ
   if (partition_horz_allowed &&
       (do_rect || vp9_active_h_edge(cpi, mi_row, mi_step))) {
-      subsize = get_subsize(bsize, PARTITION_HORZ);
+    subsize = get_subsize(bsize, PARTITION_HORZ);
     if (cpi->sf.adaptive_motion_search)
       load_pred_mv(x, ctx);
     if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
@@ -2673,6 +2686,10 @@ static void rd_pick_partition(VP9_COMP *cpi, ThreadData *td,
       if (sum_rdc.rdcost < best_rdc.rdcost) {
         best_rdc = sum_rdc;
         pc_tree->partitioning = PARTITION_HORZ;
+
+        if ((cpi->sf.less_rectangular_check) &&
+            (bsize > cpi->sf.use_square_only_threshold))
+          do_rect = 0;
       }
     }
     restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
@@ -2680,7 +2697,7 @@ static void rd_pick_partition(VP9_COMP *cpi, ThreadData *td,
   // PARTITION_VERT
   if (partition_vert_allowed &&
       (do_rect || vp9_active_v_edge(cpi, mi_col, mi_step))) {
-      subsize = get_subsize(bsize, PARTITION_VERT);
+    subsize = get_subsize(bsize, PARTITION_VERT);
 
     if (cpi->sf.adaptive_motion_search)
       load_pred_mv(x, ctx);
@@ -2734,7 +2751,6 @@ static void rd_pick_partition(VP9_COMP *cpi, ThreadData *td,
   (void) best_rd;
   *rd_cost = best_rdc;
 
-
   if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX &&
       pc_tree->index != 3) {
     int output_enabled = (bsize == BLOCK_64X64);

vp9/encoder/vp9_speed_features.c

@@ -113,8 +113,14 @@ static void set_good_speed_feature(VP9_COMP *cpi, VP9_COMMON *cm,
                                    SPEED_FEATURES *sf, int speed) {
   const int boosted = frame_is_boosted(cpi);
 
+  sf->partition_search_breakout_dist_thr = (1 << 20);
+  sf->partition_search_breakout_rate_thr = 80;
+  sf->tx_size_search_breakout = 1;
   sf->adaptive_rd_thresh = 1;
   sf->allow_skip_recode = 1;
+  sf->less_rectangular_check = 1;
+  sf->use_square_partition_only = !frame_is_boosted(cpi);
+  sf->use_square_only_threshold = BLOCK_16X16;
 
   if (speed >= 1) {
     if ((cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) ||
@@ -123,6 +129,7 @@ static void set_good_speed_feature(VP9_COMP *cpi, VP9_COMMON *cm,
     } else {
       sf->use_square_partition_only = !frame_is_intra_only(cm);
     }
+    sf->use_square_only_threshold = BLOCK_4X4;
 
     sf->less_rectangular_check  = 1;
 
@@ -139,9 +146,6 @@ static void set_good_speed_feature(VP9_COMP *cpi, VP9_COMMON *cm,
     sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V;
     sf->intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V;
     sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V;
-
-    sf->tx_size_search_breakout = 1;
-    sf->partition_search_breakout_rate_thr = 80;
   }
 
   if (speed >= 2) {
@@ -471,6 +475,7 @@ void vp9_set_speed_features_framesize_independent(VP9_COMP *cpi) {
   sf->partition_search_type = SEARCH_PARTITION;
   sf->less_rectangular_check = 0;
   sf->use_square_partition_only = 0;
+  sf->use_square_only_threshold = BLOCK_SIZES;
   sf->auto_min_max_partition_size = NOT_IN_USE;
   sf->rd_auto_partition_min_limit = BLOCK_4X4;
   sf->default_max_partition_size = BLOCK_64X64;

vp9/encoder/vp9_speed_features.h

@@ -267,6 +267,7 @@ typedef struct SPEED_FEATURES {
 
   // Disable testing non square partitions. (eg 16x32)
   int use_square_partition_only;
+  BLOCK_SIZE use_square_only_threshold;
 
   // Sets min and max partition sizes for this 64x64 region based on the
   // same 64x64 in last encoded frame, and the left and above neighbor.